Part of overall "improve filtering situation" (#116980): replace Subsampled3x3
(added for blender 3.5 in f210842a72 et al.) strip scaling filter with a
general Box filter.
Subsampled3x3 is really a Box filter ("average pixel values over NxM region"),
hardcoded to 3x3 size. As such, it works pretty well when downscaling images by
3x on each axis. But starts to break down and introduce aliasing at other
scaling factors. Also when scaling up or scaling down by less than 3x, using
total of 9 samples is a bit of overkill and hurts performance.
So instead, calculate the amount of NxM samples needed by looking at scaling
factors on X/Y axes. Note: use at least 2 samples on each axis, so that when
rotation is present, the result edges will get some anti-aliasing, just like it
was happening in previous filter implementation.
Images in PR.
Pull Request: https://projects.blender.org/blender/blender/pulls/117584
Part of overall "improve filtering situation" (#116980) task:
Add "Cubic Mitchell" filtering option to VSE strips. This is a cubic (4x4)
filter that generally looks better than bilinear, while not blurring the image
as much as the Cubic BSpline filter that exists elsewhere within Blender. It is
also default in many other apps.
Rename the (very recently added) VSE Bicubic filter option to Cubic BSpline.
Images in the PR.
Pull Request: https://projects.blender.org/blender/blender/pulls/117517
Make Subsampling 3x3 filter twice faster (on 4K UHD resolution,
Windows/VS2022/Ryzen5950X: 52.7ms -> 28.3ms), by reformulating how it works:
Conceptually Subsampling filter is a box filter: it sums up N source image
pixels, computes their average and outputs the result. Critical thing is,
that should be done in premultiplied space so that colors from fully or
mostly transparent regions do not "override" opaque colors.
Previously, when operating on byte images, the code achieved this by always
working on byte values, doing "progressively smaller" lerp into byte color
result, taking care of premultiplication and again storing the "straight"
alpha for each sample being processed. This meant that for each sample, there
are 3 divisions involved! This also led to some precision loss, since for all
9 samples all the intermediate results would only be stored at byte precision.
Reformulate that by simply accumulating the premultiplied color as a float.
This gets rid of all divisions, except the last step when said float needs to
be written back into a byte color.
The unit test results have a tiny difference, since now it is arguably better
(as per above, previously it was having some precision loss).
Pull Request: https://projects.blender.org/blender/blender/pulls/117125
Double precision pixel coordinate interpolation was added in 3a65d2f591 to
fix an issue of "wobbly" resulting image at very high scale factors. But
the root cause of that was the fact that the scanline loop was repeatedly
adding the floating point step for each pixel, instead of doing multiplication
by pixel index (repeated floating point additions can "drift" due to
imprecision, whereas multiplications are much more accurate).
Change all that math back to use single precision floats. I checked the
original issue the commit was fixing, it is still fine. Also added a gtest
to cover this situation: `imbuf_transform, nearest_very_large_scale`
This makes `IMB_transform` a tiny bit faster, on Windows/VS2022/Ryzen5950X
scaling an image at 4K resolution:
- Bilinear: 5.92 -> 5.83 ms
- Subsampled3x3: 53.6 -> 52.7 ms
Pull Request: https://projects.blender.org/blender/blender/pulls/117160
Part of overall "improve filtering situation" (#116980) task:
* Add Bicubic filtering option to strip Transform "Filter" setting.
Previously this option only existed in Transform Effect "Interpolation"
setting.
- With this addition, it feels like the transform effect could
possibly be marked as legacy/deprecated, since the regular Transform
that is on all strips can do everything that Transform Effect did?
* Speed up bicubic filtering (used now in VSE, but also in CPU Compositor,
image paint, etc.) by slightly simplifying the code and using some SIMD.
Upscaling 96x54 image to 3840x2160 resolution, using Bicubic filtering:
- Windows (VS2022, Ryzen 5950X): 35.5ms -> 15.1ms
- Mac (clang 15, M1 Max): 29.6ms -> 24.4ms
* Add gtest coverage for bicubic functionality.
Pull Request: https://projects.blender.org/blender/blender/pulls/117100
